rng_mrg gpu implementation does not support more than (2^31 -1) samples
dbarg opened this issue · comments
Hello,
Is it possible to get a fix for this?
Is it as simple as int32 -> int64 (or if not a config setting for int32->int64)?
I moved from tensorflow to theano precisely to reduce memory enough to fit on 12 GB GPUs (which worked) but now I'm hitting this limit (and I still have free memory left).
Thanks!
Error follows:
ValueError: rng_mrg gpu implementation does not support more than (2**31 -1) samples
Apply node that caused the error: GPUA_mrg_uniform{GpuArrayType(float16, matrix),inplace}(<GpuArrayType(int32, matrix)>, TensorConstant{(2,) of 63500})
Toposort index: 0
Inputs types: [GpuArrayType(int32, matrix), TensorType(int64, vector)]
Inputs shapes: [(15360, 6), (2,)]
Inputs strides: [(24, 4), (8,)]
Inputs values: ['not shown', array([63500, 63500])]
Inputs type_num: [5, 7]
Outputs clients: [['output'], [GpuElemwise{Composite{(i0 + (i1 * i2))}}[(0, 2)](GpuArrayConstant{[[-0.006874]]}, GpuArrayConstant{[[0.01375]]}, GPUA_mrg_uniform{GpuArrayType(float16, matrix),inplace}.1)]]
Debugprint of the apply node:
GPUA_mrg_uniform{GpuArrayType(float16, matrix),inplace}.0 [id A] <GpuArrayType(int32, matrix)> ''
|<GpuArrayType(int32, matrix)> [id B] <GpuArrayType(int32, matrix)>
|TensorConstant{(2,) of 63500} [id C] <TensorType(int64, vector)>
GPUA_mrg_uniform{GpuArrayType(float16, matrix),inplace}.1 [id A] <GpuArrayType(float16, matrix)> ''